def df_squeezeSeg_config():
"""Specify the parameters to tune below."""
mc = base_model_config('DF')
mc.BATCH_SIZE = 16
mc.AZIMUTH_LEVEL = 1024
mc.ZENITH_LEVEL = 32
mc.LCN_HEIGHT = 3
mc.LCN_WIDTH = 5
mc.RCRF_ITER = 3
mc.BILATERAL_THETA_A = np.array([.9, .9, .6, .6])
mc.BILATERAL_THETA_R = np.array([.015, .015, .01, .01])
mc.BI_FILTER_COEF = 0.1
mc.ANG_THETA_A = np.array([.9, .9, .6, .6])
mc.ANG_FILTER_COEF = 0.02
mc.CLS_LOSS_COEF = 15.0
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 8000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.DATA_AUGMENTATION = True
mc.RANDOM_FLIPPING = True
# x, y, z, intensity, distance
mc.INPUT_MEAN = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
mc.INPUT_STD = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
return mc
def imageNet_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
mc.IMAGE_WIDTH = 224
mc.IMAGE_HEIGHT = 224
mc.BATCH_SIZE = 256
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.QUANTIZATION = False
mc.IMAGENET = True
return mc
def kitti_model_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
# mc.IMAGE_WIDTH = 1864 # half width 621
# mc.IMAGE_HEIGHT = 562 # half height 187
mc.IMAGE_WIDTH = 1242 # half width 621
mc.IMAGE_HEIGHT = 375 # half height 187
# mc.IMAGE_WIDTH = 621
# mc.IMAGE_HEIGHT = 187
mc.WEIGHT_DECAY = 0.0001
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.LEARNING_RATE = 0.01
mc.MOMENTUM = 0.9
mc.DECAY_STEPS = 10000
mc.LR_DECAY_FACTOR = 0.5
mc.BATCH_SIZE = 20
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.MAX_GRAD_NORM = 1.0
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
mc.USE_DECONV = False
mc.EXCLUDE_HARD_EXAMPLES = False
return mc
def imagenet_config():
"""Specify the parameters to tune below."""
mc = base_model_config('ILSVRC2013')
mc.DEBUG_MODE = False
# Data Augmentation
mc.DATA_AUGMENTATION = True
# Network Architecture
mc.BN = True
mc.LOAD_BN = True
# Remove BGR mean
mc.SUB_BGR_MEANS = True
mc.IMAGE_WIDTH = 224
mc.IMAGE_HEIGHT = 224
mc.BATCH_SIZE = 32
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 1e-3
mc.LR_POLICY = 'step'
mc.LR_STEP_BOUNDRY = [10000, 15000]
mc.LR_STEP_VALUE = [1e-3, 1e-4, 1e-5]
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
return mc
def cityscapes_squeezeDet_config():
"""Specify the parameters to tune below."""
mc = base_model_config('cityscapes')
mc.IMAGE_WIDTH = 320 # 1248
mc.IMAGE_HEIGHT = 160 # 384
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 38 # 150
mc.DRIFT_Y = 41 # 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
return mc
def kitti_vgg16_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
mc.IMAGE_WIDTH = 224
mc.IMAGE_HEIGHT = 224
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_SHAPE = np.array(
[[100, 100]])
mc.ANCHOR_BOX = set_anchors(mc, 1)
mc.ANCHOR_BOX2 = set_anchors(mc, 2)
mc.ANCHOR_BOX3 = set_anchors(mc, 4)
# print(mc.ANCHOR_BOX3)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHORS2 = len(mc.ANCHOR_BOX2)
mc.ANCHORS3 = len(mc.ANCHOR_BOX3)
# print(mc.ANCHORS, mc.ANCHORS2, mc.ANCHORS3)
mc.ANCHOR_TOTAL = mc.ANCHORS + mc.ANCHORS2 + mc.ANCHORS3
mc.ANCHOR_PER_GRID = 1
# mc.ANCHOR_SHAPE = np.array(
# [[ 36., 37.], [ 366., 174.], [ 115., 59.],
# [ 162., 87.], [ 38., 90.], [ 258., 173.],
# [ 224., 108.], [ 78., 170.], [ 72., 43.]])
mc.PYRAMID_SCALE = [1, 2, 4]
mc.SCALE_SIZE = 3
mc.LOAD_PRETRAINED_MODEL = True
return mc
def kitti_squeezeSeg32_config_ext():
""" Parameters to tune after adding one class """
mc = base_model_config('KITTI')
mc.CLASSES = ['unknown', 'car', 'pedestrian', 'cyclist', 'ground']
mc.NUM_CLASS = len(mc.CLASSES)
mc.CLS_2_ID = dict(zip(mc.CLASSES, range(len(mc.CLASSES))))
#mc.CLS_LOSS_WEIGHT = np.array([1/15.0, 1.0, 10.0, 10.0, 1/10.0])
mc.CLS_LOSS_WEIGHT = np.array([0.00399712, 1.0, 0.68579656, 0.68579656, 0.01329])
mc.CLS_COLOR_MAP = np.array([[ 0.00, 0.00, 0.00],
[ 0.12, 0.56, 0.37],
[ 0.66, 0.55, 0.71],
[ 0.58, 0.72, 0.88],
[ 0.25, 0.51, 0.76]])
mc.BATCH_SIZE = 32
mc.AZIMUTH_LEVEL = 512
mc.ZENITH_LEVEL = 32
mc.LCN_HEIGHT = 3
mc.LCN_WIDTH = 5
#mc.RCRF_ITER = 3
mc.RCRF_ITER = 5
# Not sure if I need to change this
mc.BI_FILTER_COEF = 0.1
mc.BILATERAL_THETA_A = np.array([0.9, 0.9, 0.6, 0.6, 0.9])
mc.BILATERAL_THETA_R = np.array([0.015, 0.015, 0.01, 0.01, 0.015])
mc.ANG_THETA_A = np.array([0.9, 0.9, 0.6, 0.6, 0.9])
mc.ANG_FILTER_COEF = 0.02
mc.CLS_LOSS_COEF = 15.0
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.DATA_AUGMENTATION = True
mc.RANDOM_FLIPPING = True
# x, y, z, intensity, distance
# mc.INPUT_MEAN = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
mc.INPUT_MEAN = np.array([[[7.11174452147, 0.0943776729418, -0.500243253951, 0.110507476635,7.70754909688]]])
# mc.INPUT_STD = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
mc.INPUT_STD = np.array([[[10.9252722964, 4.8862697957, 0.775360202555, 0.166859600962,11.6589714541]]])
mc.num_of_input_channels=5
mc.use_focal_loss=False
mc.EVAL_ON_ORG=False
return mc
def kitti_zynqDet_FPN_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
# mc.IMAGE_WIDTH = 1242
# mc.IMAGE_HEIGHT = 375
mc.IMAGE_WIDTH = 640
mc.IMAGE_HEIGHT = 480
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
# anchor_shapes = np.reshape(
# [np.array(
# [[ 127., 100.], [ 115., 162.], [ 248., 166.],
# [ 161., 221.], [ 173., 150.], [ 191., 94.],
# [ 85., 211.], [ 67., 118.], [ 70., 161.]])] * H * W,
# (H, W, B, 2)
# )
mc.ANCHOR_SHAPES = np.array([[224, 224]])
mc.ANCHOR_PER_GRID = 1
mc.ANCHOR_BOX = set_anchors(mc, 1)
mc.ANCHOR_BOX2 = set_anchors(mc, 2)
mc.ANCHOR_BOX3 = set_anchors(mc, 4)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHORS2 = len(mc.ANCHOR_BOX2)
mc.ANCHORS3 = len(mc.ANCHOR_BOX3)
mc.ANCHOR_TOTAL = mc.ANCHORS + mc.ANCHORS2 + mc.ANCHORS3
return mc
def kitti_squeezeSeg_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
print("Kitti preloaded =", mc.LOAD_PRETRAINED_MODEL)
print("Loaded model path = ", mc.PRETRAINED_MODEL_PATH)
mc.CLASSES = ['background', 'car']
mc.NUM_CLASS = len(mc.CLASSES)
mc.CLS_2_ID = dict(zip(mc.CLASSES, range(len(mc.CLASSES))))
# controlling the relative importance of weights
mc.CLS_LOSS_WEIGHT = np.array([1.0 / 15.0, 1.0])
mc.CLS_COLOR_MAP = np.array([[0.00, 0.00, 0.00], [0.12, 0.56, 0.37]])
mc.BATCH_SIZE = 32
# how many pixel long our projection is
mc.AZIMUTH_LEVEL = 512
# how many pixel tall our projection is
mc.ZENITH_LEVEL = 64
# Conditional Random Field
mc.LCN_HEIGHT = 3
mc.LCN_WIDTH = 5
# How many times it is passed through the RNN
mc.RCRF_ITER = 3
mc.BILATERAL_THETA_A = np.array([.9, .9, .6, .6])
mc.BILATERAL_THETA_R = np.array([.015, .015, .01, .01])
mc.BI_FILTER_COEF = 0.1
mc.ANG_THETA_A = np.array([.9, .9, .6, .6])
mc.ANG_FILTER_COEF = 0.02
# Loss coefficient
mc.CLS_LOSS_COEF = 15.0
mc.WEIGHT_DECAY = 0.00001
mc.LEARNING_RATE = 0.0001
mc.DECAY_STEPS = 100
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.1
mc.DATA_AUGMENTATION = True
mc.RANDOM_FLIPPING = True
#mc.INPUT_MEAN = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
#mc.INPUT_STD = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
# x, y, z, intensity, distance
mc.INPUT_MEAN = np.array([[[0, 0, 5, 0.5, 6]]])
mc.INPUT_STD = np.array([[[1, 1, 1, 1, 1]]])
return mc
def kitti_squeezeSeg16_config_ext():
""" Parameters to tune after adding one class """
mc = base_model_config('KITTI')
mc.CLASSES = ['unknown', 'car', 'pedestrian', 'cyclist', 'ground']
mc.NUM_CLASS = len(mc.CLASSES)
mc.CLS_2_ID = dict(zip(mc.CLASSES, range(len(mc.CLASSES))))
mc.CLS_LOSS_WEIGHT = np.array([1/15.0, 1.0, 10.0, 10.0, 1/15.0])
mc.CLS_COLOR_MAP = np.array([[ 0.00, 0.00, 0.00],
[ 0.12, 0.56, 0.37],
[ 0.66, 0.55, 0.71],
[ 0.58, 0.72, 0.88],
[ 0.25, 0.51, 0.76]])
mc.BATCH_SIZE = 32
mc.AZIMUTH_LEVEL = 512
mc.ZENITH_LEVEL = 16
mc.LCN_HEIGHT = 3
mc.LCN_WIDTH = 5
mc.RCRF_ITER = 3
# Not sure if I need to change this
mc.BI_FILTER_COEF = 0.1
mc.BILATERAL_THETA_A = np.array([.9, .9, .6, .6, .9])
mc.BILATERAL_THETA_R = np.array([.015, .015, .01, .01, .015])
mc.ANG_THETA_A = np.array([.9, .9, .6, .6, .9])
mc.ANG_FILTER_COEF = 0.02
mc.CLS_LOSS_COEF = 15.0
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.DATA_AUGMENTATION = True
mc.RANDOM_FLIPPING = True
# x, y, z, intensity, distance
mc.INPUT_MEAN = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
mc.INPUT_STD = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
mc.num_of_input_channels=5
mc.use_focal_loss=False
mc.EVAL_ON_ORG=False
return mc
def pascal_voc_yolo_config():
"""Specify the parameters to tune below."""
mc = base_model_config('PASCAL_VOC')
mc.DEBUG_MODE = True
mc.SUB_BGR_MEANS = False
# Data Augmentation
mc.LOSS_TYPE = 'YOLO'
mc.DATA_AUG_TYPE = 'YOLO'
# Network Architecture
mc.BN = True
mc.LOAD_BN = True
mc.IMAGE_WIDTH = 416
mc.IMAGE_HEIGHT = 416
mc.BATCH_SIZE = 32
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 1e-3
mc.LR_POLICY = 'step'
mc.LR_STEP_BOUNDRY = [10000, 15000]
mc.LR_STEP_VALUE = [1e-3, 1e-4, 1e-5]
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LOSS_COEF_BBOX = 1.0
mc.LOSS_COEF_CONF_POS = 5.0
mc.LOSS_COEF_CONF_NEG = 1.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.4
mc.DATA_AUGMENTATION = True
mc.NET_OUT_SHAPE = (13, 13, 5) # (H, W, B)
mc.ANCHOR_BOX = set_anchors(mc, mc.NET_OUT_SHAPE[-1])
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 5
return mc
def kitti_squeezeDet_FPN_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
# mc.IMAGE_WIDTH = 1242
# mc.IMAGE_HEIGHT = 375
mc.IMAGE_WIDTH = 224
mc.IMAGE_HEIGHT = 224
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_SHAPES = np.array([[100, 100]])
mc.ANCHOR_PER_GRID = 1
mc.ANCHOR_BOX = set_anchors(mc, 1)
mc.ANCHOR_BOX2 = set_anchors(mc, 2)
mc.ANCHOR_BOX3 = set_anchors(mc, 4)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHORS2 = len(mc.ANCHOR_BOX2)
mc.ANCHORS3 = len(mc.ANCHOR_BOX3)
mc.ANCHOR_TOTAL = mc.ANCHORS + mc.ANCHORS2 + mc.ANCHORS3
return mc
def pascal_voc_vgg16_config():
"""Specify the parameters to tune below."""
mc = base_model_config('PASCAL_VOC')
mc.DEBUG_MODE = False
# Data Augmentation
#mc.LOSS_TYPE = 'YOLO'
mc.DATA_AUG_TYPE = 'YOLO'
# Network Architecture
mc.BN = True
mc.IMAGE_WIDTH = 416
mc.IMAGE_HEIGHT = 416
mc.BATCH_SIZE = 32
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 1e-3
mc.DECAY_STEPS = 2e4
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 50
mc.DRIFT_Y = 50
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
return mc
def kitti_squeezeSeg_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
mc.CLASSES = ['unknown', 'car', 'pedestrian', 'cyclist']
mc.NUM_CLASS = len(mc.CLASSES)
mc.CLS_2_ID = dict(zip(mc.CLASSES, range(len(mc.CLASSES))))
mc.CLS_LOSS_WEIGHT = np.array([1/3.0, 1.0, 3.5, 3.5])
mc.CLS_COLOR_MAP = np.array([[ 0.00, 0.00, 0.00],
[ 0.12, 0.56, 0.37],
[ 0.66, 0.55, 0.71],
[ 0.58, 0.72, 0.88]])
mc.BATCH_SIZE = 6
mc.AZIMUTH_LEVEL = 512
mc.ZENITH_LEVEL = 64
mc.FOCAL_GAMMA = 2.0
mc.LCN_HEIGHT = 3
mc.LCN_WIDTH = 5
mc.RCRF_ITER = 3
mc.BILATERAL_THETA_A = np.array([.9, .9, .6, .6])
mc.BILATERAL_THETA_R = np.array([.015, .015, .01, .01])
mc.BI_FILTER_COEF = 0.1
mc.ANG_THETA_A = np.array([.9, .9, .6, .6])
mc.ANG_FILTER_COEF = 0.02
mc.CLS_LOSS_COEF = 15.0
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.05
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.DATA_AUGMENTATION = True
mc.RANDOM_FLIPPING = True
# x, y, z, intensity, distance
mc.INPUT_MEAN = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
mc.INPUT_STD = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
return mc
def vid_yolo_config():
"""Specify the parameters to tune below."""
mc = base_model_config('VID')
mc.DEBUG_MODE = False
# Data Augmentation
mc.LOSS_TYPE = 'YOLO'
mc.DATA_AUG_TYPE = 'YOLO'
# Network Architecture
mc.BN = True
mc.IMAGE_WIDTH = 416
mc.IMAGE_HEIGHT = 416
mc.BATCH_SIZE = 64
mc.WEIGHT_DECAY = 1e-4
mc.LEARNING_RATE = 1e-3
mc.DECAY_STEPS = 2e4
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 1.0
mc.LOSS_COEF_CONF_POS = 5.0
mc.LOSS_COEF_CONF_NEG = 1.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
return mc
def kitti_shuffledet_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
mc.IMAGE_WIDTH = 1248
mc.IMAGE_HEIGHT = 384
# mc.IMAGE_WIDTH = 560
# mc.IMAGE_HEIGHT = 180
mc.BATCH_SIZE = 8
mc.WEIGHT_DECAY = 0.0005
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
# mc.DRIFT_X = 150/(1248./560)
# mc.DRIFT_Y = 100/(384./180)
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
return mc
def kitti_res50_filter_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
mc.IMAGE_WIDTH = 621
mc.IMAGE_HEIGHT = 187
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
net_output_height, net_output_width = 24, 78
mc.GRID_CELLS = net_output_height * net_output_width
mc.CLASS_BIASSES = np.array([1.0, 1.0, 1.0], dtype=np.float32) / 3
return mc
def kitti_squeezeSeg_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
mc.NUM_BIN = 10
mc.BIN_INTERVALS = np.array([100] * 10)
mc.MID_VALUES = np.array(range(50, 1000, 100))
mc.BIN_LOSS_WEIGHT = np.array([5, 5, 4, 4, 3, 3, 2, 2, 1, 1])
mc.BIN_LOSS_COEF = 4
mc.DELTA_LOSS_COEF = 30e-4
mc.SOFT_LABEL_SIGMA = 0.5
mc.BATCH_SIZE = 48
mc.AZIMUTH_LEVEL = 512
mc.ZENITH_LEVEL = 64
mc.LCN_HEIGHT = 3
mc.LCN_WIDTH = 5
mc.RCRF_ITER = 3
mc.BI_FILTER_COEF = 0.1
mc.ANG_FILTER_COEF = 0.02
mc.ECULIDEAN_LOSS_COEF = 15.0
mc.CLS_LOSS_COEF = 15.0
mc.WEIGHT_DECAY = 0.0001
#mc.LEARNING_RATE = 0.01
mc.LEARNING_RATE = 0.05
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.DATA_AUGMENTATION = True
mc.RANDOM_FLIPPING = True
# x, y, z, intensity, distance
mc.INPUT_MEAN_KITTI = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
mc.INPUT_STD_KITTI = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
mc.INPUT_MEAN_GTAV = np.array([[[7.98, 0.22, -0.67, 8.91]]])
mc.INPUT_STD_GTAV = np.array([[[9.82, 5.43, 0.73, 10.56]]])
return mc
def kitti_squeezeDet_config():
"""Specify the parameters to tune below."""
mc = base_model_config('KITTI')
# mc.IMAGE_WIDTH = 1248
# mc.IMAGE_HEIGHT = 384
mc.IMAGE_WIDTH = 624 # 2448 / 4 -> 39 x 16
mc.IMAGE_HEIGHT = 512 # 2048 / 4
# mc.IMAGE_WIDTH = 1232 # 2448 / 2 -> 77 x 16
# mc.IMAGE_HEIGHT = 1024 # 2048 / 2
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 10000
mc.MAX_GRAD_NORM = 1.0
mc.MOMENTUM = 0.9
mc.LR_DECAY_FACTOR = 0.5
mc.LOSS_COEF_BBOX = 5.0
mc.LOSS_COEF_CONF_POS = 75.0
mc.LOSS_COEF_CONF_NEG = 100.0
mc.LOSS_COEF_CLASS = 1.0
mc.PLOT_PROB_THRESH = 0.4
mc.NMS_THRESH = 0.4
mc.PROB_THRESH = 0.005
mc.TOP_N_DETECTION = 64
mc.DATA_AUGMENTATION = True
mc.DRIFT_X = 150
mc.DRIFT_Y = 100
mc.EXCLUDE_HARD_EXAMPLES = False
mc.ANCHOR_BOX = set_anchors(mc)
mc.ANCHORS = len(mc.ANCHOR_BOX)
mc.ANCHOR_PER_GRID = 9
return mc
def alibaba_squeezeSeg_config():
asc = base_model_config('KITTI')
asc.CLASSES = [
'DontCare',
'cyclist',
'tricycle',
'sm_allMot',
'bigMot',
'pedestrian',
'crowds',
'unknown',
]
asc.NUM_CLASS = len(asc.CLASSES)
asc.CLS_2_ID = dict(zip(asc.CLASSES, range(len(asc.CLASSES))))
asc.CLS_LOSS_WEIGHT = np.array(
[1 / 10.0, 1.0, 10.0, 11.0, 8.0, 7.0, 9.0, 20.0])
asc.CLS_COLOR_MAP = np.array([[1.0, 1.0, 1.0], [1.0, 0.0, 0.0],
[0.0, 1.0, 0.0], [0.0, 0.0, 1.0],
[1.0, 1.0, 0.0], [1.0, 0.0, 1.0],
[0.0, 1.0, 1.0], [1.0, 0.5, 0.0]])
asc.BATCH_SIZE = 32
asc.AZIMUTH_LEVEL = 512
asc.ZENITH_LEVEL = 64
asc.LCN_HEIGHT = 3
asc.LCN_WIDTH = 5
asc.RCRF_ITER = 3
asc.BILATERAL_THETA_A = np.array([.9, .6, .9, .6, .9, .6, .9, .6])
asc.BILATERAL_THETA_R = np.array(
[.015, .01, .015, .01, .015, .01, .015, .01])
asc.BI_FILTER_COEF = 0.1
asc.ANG_THETA_A = np.array([.9, .6, .9, .6, .9, .6, .9, .6])
asc.ANG_FILTER_COEF = 0.02
asc.CLS_LOSS_COEF = 15.0
asc.WEIGHT_DECAY = 0.0008 # 0.0001
asc.LEARNING_RATE = 0.01 # origin 0.01
asc.DECAY_STEPS = 10000
asc.MAX_GRAD_NORM = 1.0
asc.MOMENTUM = 0.9
asc.LR_DECAY_FACTOR = 0.5
asc.DATA_AUGMENTATION = False
asc.RANDOM_FLIPPING = False
# x, y, z, intensity, distance
asc.INPUT_MEAN = np.array([[[10.88, 0.23, -1.04, 0.21, 12.12]]])
asc.INPUT_STD = np.array([[[11.47, 6.91, 0.86, 0.16, 12.32]]])
return asc
def vkitti_SSD_config():
"""Specify the parameters to tune below."""
mc = base_model_config('VKITTI')
mc.VERSION = 'V1'
mc.IMAGE_WIDTH = 300
mc.IMAGE_HEIGHT = 300
mc.BGR_MEANS = np.array([[[104,117,123]]])
mc.BATCH_SIZE = 20
mc.WEIGHT_DECAY = 0.0005 #0.00001
mc.LEARNING_RATE = 0.01
mc.DECAY_STEPS = 20000
mc.LR_DECAY_FACTOR = 0.5
mc.MAX_GRAD_NORM = 1.0 #1.0
mc.MOMENTUM = 0.9
mc.mbox_source_layers_name = ['conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2']
mc.normalizations = [20, -1, -1, -1, -1, -1]
mc.bQuant = False
cal_prior_param(mc)
mc.BACKGROUD_ID = 0
mc.TOP_N_DETECTION = 400
mc.PROB_THRESH = 0.01
mc.NMS_THRESH = 0.45
mc.PLOT_PROB_THRESH = 0.4
mc.overlap_threshold = 0.5
#mc.keep_top_k = 200
print ('vkitti_SSD_config')
#data layer params
mc.batch_sampler = get_batch_sampler()
mc.expand_param = edict()
mc.expand_param.prob = 1.0
mc.expand_param.min_expand_ratio = 1.0
mc.expand_param.max_expand_ratio = 3.0
#multibox_loss layer params
mc.multibox_loss_param = edict()
mc.multibox_loss_param.overlap_threshold = 0.5
mc.multibox_loss_param.neg_pos_ratio = 3.
mc.multibox_loss_param.neg_overlap = 0.5
mc.background_label_id = 0
#anchors
mc.ANCHOR_BOX,mc.ANCHORS_NUM,mc.ANCHOR_PER_GRID = all_anchors(mc)
mc.PRIORBOX_VARIANCES = all_prior_variance(mc)
'''
print ('np.shape(mc.ANCHOR_BOX):',np.shape(mc.ANCHOR_BOX))
print ('mc.ANCHORS_NUM:',mc.ANCHORS_NUM)
print ('np.shape(mc.PRIORBOX_VARIANCES):',np.shape(mc.PRIORBOX_VARIANCES))
priorboxes = np.concatenate([mc.ANCHOR_BOX,mc.PRIORBOX_VARIANCES],axis=0)
save_array_to_txt_file('all_anchors.txt',mc.ANCHOR_BOX)
save_array_to_txt_file('all_prior_variance.txt',mc.PRIORBOX_VARIANCES)
save_array_to_txt_file('_priorboxes.txt',priorboxes)
raw_input('pause')
'''
mc.ADD_WEIGHT_DECAY_TO_LOSS = True
mc.NUM_THREAD = 4
mc.QUEUE_CAPACITY = 100
return mc